We present a computational model of false recall phenomena. The model is based on an integrated architecture including modules that have been successful at accounting for other types of memory tasks. Words presented in a list are represented by semantic vectors from a co-occurrence model, and are encoded into a composite store. At test, the model generates a list of candidate words from the lexicon and decides which of these words to recall using a recognition process. We show that the model is able to account for a wide range of effects in false recall, including levels of false recall in DRM studies, item-level effects, number of associates, and categorical vs. associative list structure.